2010-09-15 18:36:07 +04:00
|
|
|
/***********************************************************************
|
|
|
|
Moses - factored phrase-based language decoder
|
|
|
|
Copyright (C) 2010 University of Edinburgh
|
|
|
|
|
|
|
|
This library is free software; you can redistribute it and/or
|
|
|
|
modify it under the terms of the GNU Lesser General Public
|
|
|
|
License as published by the Free Software Foundation; either
|
|
|
|
version 2.1 of the License, or (at your option) any later version.
|
|
|
|
|
|
|
|
This library is distributed in the hope that it will be useful,
|
|
|
|
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
|
Lesser General Public License for more details.
|
|
|
|
|
|
|
|
You should have received a copy of the GNU Lesser General Public
|
|
|
|
License along with this library; if not, write to the Free Software
|
|
|
|
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
|
|
***********************************************************************/
|
|
|
|
|
2010-09-28 19:13:50 +04:00
|
|
|
#include <algorithm>
|
2010-09-15 19:38:46 +04:00
|
|
|
#include <cstdlib>
|
|
|
|
#include <ctime>
|
2010-09-15 18:36:07 +04:00
|
|
|
#include <string>
|
|
|
|
#include <vector>
|
|
|
|
|
|
|
|
#include <boost/program_options.hpp>
|
2010-09-28 19:13:50 +04:00
|
|
|
#ifdef MPI_ENABLE
|
|
|
|
#include <boost/mpi.hpp>
|
|
|
|
namespace mpi = boost::mpi;
|
|
|
|
#endif
|
2010-09-15 18:36:07 +04:00
|
|
|
|
2010-09-15 19:38:46 +04:00
|
|
|
#include "FeatureVector.h"
|
2010-09-15 18:36:07 +04:00
|
|
|
#include "StaticData.h"
|
2010-09-16 20:23:52 +04:00
|
|
|
#include "ChartTrellisPathList.h"
|
2010-09-17 11:35:31 +04:00
|
|
|
#include "ChartTrellisPath.h"
|
|
|
|
#include "ScoreComponentCollection.h"
|
2010-09-15 18:36:07 +04:00
|
|
|
#include "Decoder.h"
|
2010-09-15 19:38:46 +04:00
|
|
|
#include "Optimiser.h"
|
2010-09-15 18:36:07 +04:00
|
|
|
|
|
|
|
using namespace Mira;
|
|
|
|
using namespace std;
|
|
|
|
using namespace Moses;
|
|
|
|
namespace po = boost::program_options;
|
|
|
|
|
2010-09-16 20:23:52 +04:00
|
|
|
void OutputNBestList(const MosesChart::TrellisPathList &nBestList, const TranslationSystem* system, long translationId);
|
|
|
|
|
2010-09-15 19:38:46 +04:00
|
|
|
bool loadSentences(const string& filename, vector<string>& sentences) {
|
|
|
|
ifstream in(filename.c_str());
|
|
|
|
if (!in) return false;
|
|
|
|
string line;
|
|
|
|
while(getline(in,line)) {
|
|
|
|
sentences.push_back(line);
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2010-09-28 19:13:50 +04:00
|
|
|
struct RandomIndex {
|
|
|
|
ptrdiff_t operator() (ptrdiff_t max) {
|
|
|
|
return static_cast<ptrdiff_t>(rand() % max);
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2010-09-15 18:36:07 +04:00
|
|
|
int main(int argc, char** argv) {
|
2010-09-28 19:13:50 +04:00
|
|
|
size_t rank = 0; size_t size = 1;
|
|
|
|
#ifdef MPI_ENABLE
|
|
|
|
mpi::environment env(argc,argv);
|
|
|
|
mpi::communicator world;
|
|
|
|
rank = world.rank();
|
|
|
|
size = world.size();
|
|
|
|
#endif
|
|
|
|
cerr << "Rank: " << rank << " Size: " << size << endl;
|
|
|
|
|
2010-09-15 18:36:07 +04:00
|
|
|
bool help;
|
|
|
|
int verbosity;
|
|
|
|
string mosesConfigFile;
|
|
|
|
string inputFile;
|
|
|
|
vector<string> referenceFiles;
|
2010-09-28 19:13:50 +04:00
|
|
|
size_t epochs;
|
|
|
|
string learner;
|
|
|
|
bool shuffle = true;
|
|
|
|
size_t mixFrequency;
|
|
|
|
size_t weightDumpFrequency;
|
2010-09-15 18:36:07 +04:00
|
|
|
po::options_description desc("Allowed options");
|
|
|
|
desc.add_options()
|
|
|
|
("help",po::value( &help )->zero_tokens()->default_value(false), "Print this help message and exit")
|
|
|
|
("config,f",po::value<string>(&mosesConfigFile),"Moses ini file")
|
|
|
|
("verbosity,v", po::value<int>(&verbosity)->default_value(0), "Verbosity level")
|
|
|
|
("input-file,i",po::value<string>(&inputFile),"Input file containing tokenised source")
|
2010-09-28 19:13:50 +04:00
|
|
|
("reference-files,r", po::value<vector<string> >(&referenceFiles), "Reference translation files for training")
|
|
|
|
("epochs,e", po::value<size_t>(&epochs)->default_value(1), "Number of epochs")
|
|
|
|
("learner,l", po::value<string>(&learner)->default_value("mira"), "Learning algorithm")
|
|
|
|
("mix-frequency", po::value<size_t>(&mixFrequency)->default_value(1), "How often per epoch to mix weights, when using mpi")
|
|
|
|
("weight-dump-frequency", po::value<size_t>(&weightDumpFrequency)->default_value(1), "How often per epoch to dump weights");
|
2010-09-15 18:36:07 +04:00
|
|
|
|
|
|
|
po::options_description cmdline_options;
|
|
|
|
cmdline_options.add(desc);
|
|
|
|
po::variables_map vm;
|
|
|
|
po::store(po::command_line_parser(argc,argv).
|
|
|
|
options(cmdline_options).run(), vm);
|
|
|
|
po::notify(vm);
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (help) {
|
2010-09-15 19:38:46 +04:00
|
|
|
std::cout << "Usage: " + string(argv[0]) + " -f mosesini-file -i input-file -r reference-file(s) [options]" << std::endl;
|
2010-09-15 18:36:07 +04:00
|
|
|
std::cout << desc << std::endl;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (mosesConfigFile.empty()) {
|
|
|
|
cerr << "Error: No moses ini file specified" << endl;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (inputFile.empty()) {
|
|
|
|
cerr << "Error: No input file specified" << endl;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!referenceFiles.size()) {
|
|
|
|
cerr << "Error: No reference files specified" << endl;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
2010-09-28 19:13:50 +04:00
|
|
|
//FIXME: Make these configurable
|
|
|
|
float miraLowerBound = 0;
|
|
|
|
float miraUpperBound = 1;
|
|
|
|
|
|
|
|
Optimiser* optimiser = NULL;
|
|
|
|
if (learner == "mira") {
|
|
|
|
cerr << "Optimising using Mira" << endl;
|
|
|
|
optimiser = new MiraOptimiser(miraLowerBound, miraUpperBound);
|
|
|
|
} else if (learner == "perceptron") {
|
|
|
|
cerr << "Optimising using Perceptron" << endl;
|
|
|
|
optimiser = new Perceptron();
|
|
|
|
} else {
|
|
|
|
cerr << "Error: Unknown optimiser: " << learner << endl;
|
|
|
|
}
|
2010-09-15 18:36:07 +04:00
|
|
|
|
|
|
|
|
|
|
|
//load input and references
|
2010-09-15 19:38:46 +04:00
|
|
|
vector<string> inputSentences;
|
|
|
|
if (!loadSentences(inputFile, inputSentences)) {
|
|
|
|
cerr << "Error: Failed to load input sentences from " << inputFile << endl;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
vector< vector<string> > referenceSentences(referenceFiles.size());
|
|
|
|
for (size_t i = 0; i < referenceFiles.size(); ++i) {
|
|
|
|
if (!loadSentences(referenceFiles[i], referenceSentences[i])) {
|
|
|
|
cerr << "Error: Failed to load reference sentences from " << referenceFiles[i] << endl;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
if (referenceSentences[i].size() != inputSentences.size()) {
|
|
|
|
cerr << "Error: Input file length (" << inputSentences.size() <<
|
|
|
|
") != (" << referenceSentences[i].size() << ") length of reference file " << i <<
|
|
|
|
endl;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
//initialise moses
|
2010-09-16 11:49:12 +04:00
|
|
|
initMoses(mosesConfigFile, verbosity);//, argc, argv);
|
2010-09-17 19:17:33 +04:00
|
|
|
MosesDecoder* decoder = new MosesDecoder(referenceSentences) ;
|
2010-09-28 19:13:50 +04:00
|
|
|
|
|
|
|
//Optionally shuffle the sentences
|
|
|
|
vector<size_t> order;
|
|
|
|
if (rank == 0) {
|
|
|
|
for (size_t i = 0; i < inputSentences.size(); ++i) {
|
|
|
|
order.push_back(i);
|
|
|
|
}
|
|
|
|
if (shuffle) {
|
|
|
|
RandomIndex rindex;
|
|
|
|
random_shuffle(order.begin(), order.end(), rindex);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef MPI_ENABLE
|
|
|
|
mpi::broadcast(world,order,0);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
//Create the shards
|
|
|
|
vector<size_t> shard;
|
|
|
|
float shardSize = (float)(order.size()) / size;
|
|
|
|
VERBOSE(1, "Shard size: " << shardSize << endl);
|
|
|
|
size_t shardStart = (size_t)(shardSize * rank);
|
|
|
|
size_t shardEnd = (size_t)(shardSize * (rank+1));
|
|
|
|
if (rank == size-1) shardEnd = order.size();
|
|
|
|
VERBOSE(1,"Rank: " << rank << " Shard start: " << shardStart << " Shard end: " << shardEnd << endl);
|
|
|
|
shard.resize(shardSize);
|
|
|
|
copy(order.begin() + shardStart, order.begin() + shardEnd, shard.begin());
|
|
|
|
|
2010-09-17 19:17:33 +04:00
|
|
|
|
2010-09-15 18:36:07 +04:00
|
|
|
|
|
|
|
//Main loop:
|
2010-09-17 16:54:58 +04:00
|
|
|
ScoreComponentCollection cumulativeWeights;
|
|
|
|
size_t modelHypoCount = 10;
|
|
|
|
size_t hopeHypoCount = 10;
|
|
|
|
size_t fearHypoCount = 10;
|
2010-09-28 19:13:50 +04:00
|
|
|
size_t iterations = 0;
|
2010-09-15 19:38:46 +04:00
|
|
|
|
2010-09-17 11:35:31 +04:00
|
|
|
|
2010-09-15 20:09:43 +04:00
|
|
|
for (size_t epoch = 0; epoch < epochs; ++epoch) {
|
2010-09-28 19:13:50 +04:00
|
|
|
//TODO: batching
|
|
|
|
size_t shardPosition = 0;
|
|
|
|
for (vector<size_t>::const_iterator sid = shard.begin();
|
|
|
|
sid != shard.end(); ++sid) {
|
|
|
|
const string& input = inputSentences[*sid];
|
|
|
|
const vector<string>& refs = referenceSentences[*sid];
|
2010-09-15 20:09:43 +04:00
|
|
|
|
2010-09-17 16:54:58 +04:00
|
|
|
vector<vector<ScoreComponentCollection > > allScores(1);
|
|
|
|
vector<vector<float> > allLosses(1);
|
2010-09-17 14:17:27 +04:00
|
|
|
|
2010-09-15 20:09:43 +04:00
|
|
|
|
2010-09-17 11:35:31 +04:00
|
|
|
// MODEL
|
2010-09-17 18:32:27 +04:00
|
|
|
decoder->getNBest(input,
|
2010-09-28 19:13:50 +04:00
|
|
|
*sid,
|
2010-09-17 21:38:25 +04:00
|
|
|
modelHypoCount,
|
2010-09-17 18:32:27 +04:00
|
|
|
0.0,
|
|
|
|
1.0,
|
|
|
|
allScores[0],
|
|
|
|
allLosses[0]);
|
|
|
|
|
2010-09-17 11:35:31 +04:00
|
|
|
|
|
|
|
// HOPE
|
2010-09-17 18:32:27 +04:00
|
|
|
size_t oraclePos = allScores.size();
|
2010-09-17 19:17:33 +04:00
|
|
|
vector<const Word*> oracle =
|
|
|
|
decoder->getNBest(input,
|
2010-09-28 19:13:50 +04:00
|
|
|
*sid,
|
2010-09-17 21:38:25 +04:00
|
|
|
modelHypoCount,
|
2010-09-17 18:32:27 +04:00
|
|
|
1.0,
|
|
|
|
1.0,
|
|
|
|
allScores[0],
|
|
|
|
allLosses[0]);
|
|
|
|
|
|
|
|
ScoreComponentCollection oracleScores = allScores[0][oraclePos];
|
|
|
|
float oracleLoss = allLosses[0][oraclePos];
|
2010-09-17 11:35:31 +04:00
|
|
|
|
|
|
|
// FEAR
|
2010-09-17 18:32:27 +04:00
|
|
|
decoder->getNBest(input,
|
2010-09-28 19:13:50 +04:00
|
|
|
*sid,
|
2010-09-17 21:38:25 +04:00
|
|
|
modelHypoCount,
|
2010-09-17 18:32:27 +04:00
|
|
|
-1.0,
|
|
|
|
1.0,
|
|
|
|
allScores[0],
|
|
|
|
allLosses[0]);
|
2010-09-17 16:54:58 +04:00
|
|
|
|
2010-09-17 20:34:48 +04:00
|
|
|
//set loss for each sentence as oracleloss - rawsentenceloss
|
2010-09-17 16:54:58 +04:00
|
|
|
for (size_t i = 0; i < allScores.size(); ++i) {
|
|
|
|
for (size_t j = 0; j < allScores[i].size(); ++j) {
|
2010-09-17 17:16:25 +04:00
|
|
|
allLosses[i][j] = oracleLoss - allLosses[i][j];
|
2010-09-17 16:54:58 +04:00
|
|
|
}
|
|
|
|
}
|
2010-09-17 17:16:25 +04:00
|
|
|
|
2010-09-16 20:23:52 +04:00
|
|
|
|
2010-09-15 20:09:43 +04:00
|
|
|
//run optimiser
|
2010-09-17 16:54:58 +04:00
|
|
|
ScoreComponentCollection mosesWeights = decoder->getWeights();
|
|
|
|
optimiser->updateWeights(mosesWeights
|
2010-09-17 13:17:05 +04:00
|
|
|
, allScores
|
|
|
|
, allLosses
|
2010-09-17 18:32:27 +04:00
|
|
|
, oracleScores);
|
2010-09-17 11:35:31 +04:00
|
|
|
|
2010-09-15 20:09:43 +04:00
|
|
|
//update moses weights
|
2010-09-28 19:13:50 +04:00
|
|
|
mosesWeights.L1Normalise();
|
2010-09-17 16:54:58 +04:00
|
|
|
decoder->setWeights(mosesWeights);
|
2010-09-17 20:34:48 +04:00
|
|
|
|
|
|
|
//history (for approx doc bleu)
|
|
|
|
decoder->updateHistory(oracle);
|
2010-09-17 16:54:58 +04:00
|
|
|
cumulativeWeights.PlusEquals(mosesWeights);
|
|
|
|
decoder->cleanup();
|
2010-09-28 19:13:50 +04:00
|
|
|
|
|
|
|
++shardPosition;
|
|
|
|
++iterations;
|
|
|
|
|
|
|
|
//mix weights?
|
|
|
|
#ifdef MPI_ENABLE
|
|
|
|
if (shardPosition % (shard.size() / mixFrequency) == 0) {
|
|
|
|
ScoreComponentCollection averageWeights;
|
|
|
|
VERBOSE(1, "Rank: " << rank << "Before mixing: " << mosesWeights << endl);
|
|
|
|
mpi::reduce(world,mosesWeights,averageWeights,SCCPlus(),0);
|
|
|
|
if (rank == 0) {
|
|
|
|
averageWeights.MultiplyEquals(1.0f/size);
|
|
|
|
VERBOSE(1, "After mixing: " << averageWeights << endl);
|
|
|
|
}
|
|
|
|
mpi::broadcast(world,averageWeights,0);
|
|
|
|
decoder->setWeights(averageWeights);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
//dump weights?
|
|
|
|
if (shardPosition % (shard.size() / weightDumpFrequency) == 0) {
|
|
|
|
ScoreComponentCollection totalWeights(cumulativeWeights);
|
|
|
|
#ifdef MPI_ENABLE
|
|
|
|
//average across processes
|
|
|
|
mpi::reduce(world,cumulativeWeights,totalWeights,SCCPlus(),0);
|
|
|
|
#endif
|
|
|
|
if (rank == 0) {
|
|
|
|
cout << "WEIGHTS " << iterations << " ";
|
|
|
|
totalWeights.L1Normalise();
|
|
|
|
cout << totalWeights << endl;
|
|
|
|
}
|
|
|
|
}
|
2010-09-15 20:09:43 +04:00
|
|
|
}
|
2010-09-15 19:38:46 +04:00
|
|
|
}
|
2010-09-15 18:36:07 +04:00
|
|
|
|
|
|
|
|
|
|
|
|
2010-09-15 19:38:46 +04:00
|
|
|
exit(0);
|
2010-09-15 18:36:07 +04:00
|
|
|
}
|
2010-09-16 20:23:52 +04:00
|
|
|
|